Resolving Difficult Referring Expressions

Some referring expressions, such as pronominal this and that, are particularly difficult to resolve automatically and, therefore, are treated minimally if at all by most reference resolution systems. Other referring expressions, such as he, she, and they, are treated by many systems but, as yet, not with sufficient accuracy. We describe a system called CROSS (CoReference for the OntoSem2 language processing System) that automatically selects which instances of difficult referring expressions it can treat with high precision and identifies their textual antecedents. The system uses readily computable heuristic evidence in a configuration-matching approach. The identification of textual antecedents represents an intermediate result toward full reference resolution, which requires semantic and pragmatic analysis, and which augments an intelligent agent’s memory. Our evaluation shows that a language problem which seems impenetrable when viewed from the current mainstream perspective of machine learning becomes more manageable using human-inspired modeling.